I got tired of typing at my AI assistant. So I gave it a voice, a microphone, and a glowing blue orb to live in. The whole thing is two files and three dependencies. I called it JARVIS, because of course I did.
Here is what it does. You open a web page. You hold down the space bar and talk. The orb in the middle of the screen listens, and one of its rings pulses in time with your voice so you can see it hearing you. You let go, and it thinks. Then it answers out loud, in a British accent, while the text of what it is saying scrolls up the page.
The part I care about most is that it starts talking before it has finished thinking. Most voice assistants wait for the entire answer to be written, then read it to you. That pause is short, but you feel it, and it makes the thing feel like a machine reading a card. Mine watches the words arrive and speaks each sentence the moment that sentence is complete. By the time it is working out its second thought, you are already hearing the first. It feels less like waiting for a reply and more like someone talking.
It is not just a chatbot with a speaker bolted on. It can actually do things — read your files, edit them, run commands, search the web. When it does, a small gold tag appears in the transcript telling you which tool it used and on what, so you are never guessing what it just touched. It also inherits whatever tools you have already set up for yourself, rather than living in its own little sandbox.
The hardest problem turned out not to be the voice or the microphone. It was that everything an AI writes is designed to be read, not heard. Bullet points, headings, code blocks, file paths, numbers like 8765 — all of that is fine on a screen and awful out loud. A speech synthesizer will happily read you a markdown asterisk. So I told the assistant, up front, that its words were going to be spoken: keep it short, keep it conversational, no lists, no formatting, say numbers as words, and do not narrate every step of what you are doing. Fixing it at the source worked far better than trying to strip the formatting out afterwards.
The orb is the other thing I would not cut. It has no buttons and no labels, but you always know where you are: calm blue when it is idle, green while it is listening, a fast nervous pulse while it is thinking, gold while it is speaking. Four colours doing the job of an entire interface.
One honest warning. I set it to make file edits without stopping to ask permission, because being interrupted by a confirmation dialog defeats the point of talking to it hands-free. That is a real trade-off, not a detail. It is fine for a toy on my own machine. I would think hard before pointing it at anything I could not afford to lose.
Top comments (0)